AC railways are rich in harmonic signatures and feature superposition of emissions from distant sources and a strong similarity to many AC distribution grids. Rolling stock is also characterised by peculiar dynamics and changes of operating conditions. An unsupervised deep autoencoder analysis is proposed with clustering based on k-means of the latent space features. Results are provided of the resulting segmentation by the various clusters over several hours of recordings. The proposed assessment provides informative inference of voltage supply patterns, characterising the spectral content hidden in raw waveforms.